What precisely does this do? Is it for getting around the safety filter?
Title.
Yea, how do you actually use it?
It's just a lora. It makes prompt adherence stronger. The original idea came from someone else trying to bypass safety filter but it's generally useful IMO.
this lora is only 268 bytes??
Does anyone gets a black screen with this lora? mine does past 0.25
this lora is only 268 bytes??
check the script included it's a single layer
Does anyone gets a black screen with this lora? mine does past 0.2
I've heard this from ComfyUI users. Haven't been able to reproduce myself. Maybe it's from FP8 checkpoints? I developed this on the official BF16 checkpoint, I haven't tested quantizations.
lower strengths should generally work fine which is why I listed examples in the readme. Main thing is if you want "obey the prompt at all costs" you continue to get returns up to 50-100x scale depending on the image, so I just set the default to 100 which is easy to think about 1.00 == +100x strength. If your prompt is already mostly there and just needs an extra nudge then ≤0.1 strength should be totally fine.
this lora is only 268 bytes??
check the script included it's a single layer
Does anyone gets a black screen with this lora? mine does past 0.2
I've heard this from ComfyUI users. Haven't been able to reproduce myself. Maybe it's from FP8 checkpoints? I developed this on the official BF16 checkpoint, I haven't tested quantizations.
im using gguf Q5_KM quant with this the lora works until 0.25 strength then black screen is a bit annoying. BF16 are too big i cant ever tested it there
Fp8 works fine in Comfy. 0.4 is usually enough, I only tested up to 0.6.
krea2_turbo_nvfp4 from comfy org, qwen3vl_4b_mxfp8, black/empty outputs with the lora